Mapping Associations - Mapping Scientific Frontiers: The Quest for Knowledge Visualization - page 111

Fig. 3.18 Alluvial map of scientific change (Reprinted from Rosvall and Bergstrom 2010 )

structure of an evolving network at a specific point of time. Each network is divided
into a number of clusters. The corresponding clusters in adjacent networks form a
sequence of views of how the same clusters evolve over time. The split and merge
of thematic patterns can be visualized as multiple streams flow smoothly over time
(See Fig. 3.18 ).
Given a sequence of network models, one can generate alluvial maps of a diverse
range of sources of data. CiteSpace provides a function for batch exporting a series
of networks to the Pajek .net format. The exported .net files can be loaded to the
alluvial map generator (See Fig. 3.19 ). Figure 3.20 shows an alluvial map generated
based on networks of co-occurring terms in publications related to research in
regenerative medicine. Top 300 most frequently occurred terms each year are used
to construct the network of that year. The iPSCs stream identified as the most
recent thread corresponds to the research that was awarded the 2012 Nobel Prize in
Medicine. We will describe what we found in a scientometric study of regenerative
medicine shortly.
For example, Fig. 3.21 shows an alluvial map of popular tweet topics as
Hurricane Sandy moved along the east coast of the US. Each ribbon-like stream
represents a topic identified in re-tweeted tweets of each hour. For example,
the purple ribbon on Sandy Tweets runs through the entire window of analysis.
Similarly, the green ribbon, labeled as East Coast, runs through the entire night
as Hurricane Sandy moved from the South to the North along the East Coast.
Figure 3.22 shows a more complex process - a lead compound optimization. Each
stream is a depiction of the thread of the search in chemical space.

3.3 Dimensionality Reduction

Principle Component Analysis (PCA) and Multidimensional Scaling (MDS) are
classic and widely used techniques for dimensionality reduction. They are simple
to implement, efficiently computable, and guaranteed to discover the true structure